THE Spontaneous Speech: Corpus and Processing Technology

Similar documents
Linking Task: Identifying authors and book titles in verbose queries

Exploiting Phrasal Lexica and Additional Morpho-syntactic Language Resources for Statistical Machine Translation with Scarce Training Data

Learning Methods in Multilingual Speech Recognition

Disambiguation of Thai Personal Name from Online News Articles

Trend Survey on Japanese Natural Language Processing Studies over the Last Decade

Evaluation of a Simultaneous Interpretation System and Analysis of Speech Log for User Experience Assessment

Target Language Preposition Selection an Experiment with Transformation-Based Learning and Aligned Bilingual Data

The Internet as a Normative Corpus: Grammar Checking with a Search Engine

A Named Entity Recognition Method using Rules Acquired from Unlabeled Data

OCR for Arabic using SIFT Descriptors With Online Failure Prediction

The Ups and Downs of Preposition Error Detection in ESL Writing

have to be modeled) or isolated words. Output of the system is a grapheme-tophoneme conversion system which takes as its input the spelling of words,

Word Segmentation of Off-line Handwritten Documents

Using dialogue context to improve parsing performance in dialogue systems

MULTILINGUAL INFORMATION ACCESS IN DIGITAL LIBRARY

EdIt: A Broad-Coverage Grammar Checker Using Pattern Grammar

Lip reading: Japanese vowel recognition by tracking temporal changes of lip shape

STUDIES WITH FABRICATED SWITCHBOARD DATA: EXPLORING SOURCES OF MODEL-DATA MISMATCH

Ensemble Technique Utilization for Indonesian Dependency Parser

BYLINE [Heng Ji, Computer Science Department, New York University,

Intra-talker Variation: Audience Design Factors Affecting Lexical Selections

Parsing of part-of-speech tagged Assamese Texts

SINGLE DOCUMENT AUTOMATIC TEXT SUMMARIZATION USING TERM FREQUENCY-INVERSE DOCUMENT FREQUENCY (TF-IDF)

The taming of the data:

Role of Pausing in Text-to-Speech Synthesis for Simultaneous Interpretation

Improved Effects of Word-Retrieval Treatments Subsequent to Addition of the Orthographic Form

AQUA: An Ontology-Driven Question Answering System

Detecting English-French Cognates Using Orthographic Edit Distance

THE ROLE OF DECISION TREES IN NATURAL LANGUAGE PROCESSING

Speech Recognition at ICSI: Broadcast News and beyond

A Minimalist Approach to Code-Switching. In the field of linguistics, the topic of bilingualism is a broad one. There are many

Learning Computational Grammars

Chunk Parsing for Base Noun Phrases using Regular Expressions. Let s first let the variable s0 be the sentence tree of the first sentence.

Defragmenting Textual Data by Leveraging the Syntactic Structure of the English Language

Mandarin Lexical Tone Recognition: The Gating Paradigm

Vocabulary Usage and Intelligibility in Learner Language

Switchboard Language Model Improvement with Conversational Data from Gigaword

The stages of event extraction

Indian Institute of Technology, Kanpur

Semi-supervised methods of text processing, and an application to medical concept extraction. Yacine Jernite Text-as-Data series September 17.

Eyebrows in French talk-in-interaction

Twitter Sentiment Classification on Sanders Data using Hybrid Approach

Constructing Parallel Corpus from Movie Subtitles

Problems of the Arabic OCR: New Attitudes

Spoken Language Parsing Using Phrase-Level Grammars and Trainable Classifiers

Iterative Cross-Training: An Algorithm for Learning from Unlabeled Web Pages

Memory-based grammatical error correction

A Syllable Based Word Recognition Model for Korean Noun Extraction

BANGLA TO ENGLISH TEXT CONVERSION USING OPENNLP TOOLS

Modeling function word errors in DNN-HMM based LVCSR systems

Cross Language Information Retrieval

METHODS FOR EXTRACTING AND CLASSIFYING PAIRS OF COGNATES AND FALSE FRIENDS

Numeracy Medium term plan: Summer Term Level 2C/2B Year 2 Level 2A/3C

What the National Curriculum requires in reading at Y5 and Y6

A Comparison of Two Text Representations for Sentiment Analysis

Methods for the Qualitative Evaluation of Lexical Association Measures

WHEN THERE IS A mismatch between the acoustic

Edexcel GCSE. Statistics 1389 Paper 1H. June Mark Scheme. Statistics Edexcel GCSE

Web as Corpus. Corpus Linguistics. Web as Corpus 1 / 1. Corpus Linguistics. Web as Corpus. web.pl 3 / 1. Sketch Engine. Corpus Linguistics

Physics 270: Experimental Physics

Rule Learning With Negation: Issues Regarding Effectiveness

Experiments with SMS Translation and Stochastic Gradient Descent in Spanish Text Author Profiling

Reducing Features to Improve Bug Prediction

Extending Place Value with Whole Numbers to 1,000,000

Modeling function word errors in DNN-HMM based LVCSR systems

Corrective Feedback and Persistent Learning for Information Extraction

The Effect of Extensive Reading on Developing the Grammatical. Accuracy of the EFL Freshmen at Al Al-Bayt University

Backwards Numbers: A Study of Place Value. Catherine Perez

2/15/13. POS Tagging Problem. Part-of-Speech Tagging. Example English Part-of-Speech Tagsets. More Details of the Problem. Typical Problem Cases

Language Acquisition Fall 2010/Winter Lexical Categories. Afra Alishahi, Heiner Drenhaus

Universiteit Leiden ICT in Business

Derivational: Inflectional: In a fit of rage the soldiers attacked them both that week, but lost the fight.

Product Feature-based Ratings foropinionsummarization of E-Commerce Feedback Comments

Formulaic Language and Fluency: ESL Teaching Applications

LEARNING A SEMANTIC PARSER FROM SPOKEN UTTERANCES. Judith Gaspers and Philipp Cimiano

Lecture 1: Machine Learning Basics

Florida Reading Endorsement Alignment Matrix Competency 1

Assessing System Agreement and Instance Difficulty in the Lexical Sample Tasks of SENSEVAL-2

A heuristic framework for pivot-based bilingual dictionary induction

Yoshida Honmachi, Sakyo-ku, Kyoto, Japan 1 Although the label set contains verb phrases, they

Eli Yamamoto, Satoshi Nakamura, Kiyohiro Shikano. Graduate School of Information Science, Nara Institute of Science & Technology

Short Text Understanding Through Lexical-Semantic Analysis

Modeling Attachment Decisions with a Probabilistic Parser: The Case of Head Final Structures

Applications of memory-based natural language processing

Specification and Evaluation of Machine Translation Toy Systems - Criteria for laboratory assignments

Advanced Grammar in Use

Author: Justyna Kowalczys Stowarzyszenie Angielski w Medycynie (PL) Feb 2015

Radius STEM Readiness TM

Language Acquisition Chart

Learning Structural Correspondences Across Different Linguistic Domains with Synchronous Neural Language Models

arxiv: v1 [cs.cl] 2 Apr 2017

11/29/2010. Statistical Parsing. Statistical Parsing. Simple PCFG for ATIS English. Syntactic Disambiguation

CEFR Overall Illustrative English Proficiency Scales

An Evaluation of POS Taggers for the CHILDES Corpus

Rule Learning with Negation: Issues Regarding Effectiveness

Word Stress and Intonation: Introduction

Class-Discriminative Weighted Distortion Measure for VQ-Based Speaker Identification

Probabilistic Latent Semantic Analysis

Enhancing Unlexicalized Parsing Performance using a Wide Coverage Lexicon, Fuzzy Tag-set Mapping, and EM-HMM-based Lexical Probabilities

The Karlsruhe Institute of Technology Translation Systems for the WMT 2011

Transcription:

382 IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 4, JULY 2004 Morphological Analysis of the Corpus of Spontaneous Japanese Kiyotaka Uchimoto, Kazuma Takaoka, Chikashi Nobata, Atsushi Yamada, Satoshi Sekine, and Hitoshi Isahara Abstract This paper describes two methods for detecting word segments and their morphological information in a Japanese spontaneous speech corpus, and describes how to tag a large spontaneous speech corpus accurately by using the two methods. The first method is used to detect any type of word segments. The second method is used when there are several definitions for word segments and their POS categories, and when one type of word segments includes another type of word segments. In this paper, we show that by using semi-automatic analysis, we achieve a precision of better than 99% for detecting and tagging short-unit words and 97% for long-unit words; the two types of words that comprise the corpus. We also show that better accuracy is achieved by using both methods than by using only the first. Index Terms Japanese spontaneous speech corpus, maximum entropy models, morphological analysis, natural language processing, unknown words. I. INTRODUCTION THE Spontaneous Speech: Corpus and Processing Technology project is sponsoring the construction of a large spontaneous Japanese speech corpus, Corpus of Spontaneous Japanese (CSJ) [1]. The CSJ is a collection of monologues and dialogues, the majority being monologues such as Academic Presentation Speech (APS) and Simulated Public Speaking (SPS). SPS contains short speeches presented specifically for the corpus by paid nonprofessional speakers. The CSJ includes transcriptions of the talks as well as audio recordings of them. One of the goals of the project is to detect two types of word segments and corresponding morphological information in the transcriptions. The two types of word segments were defined by the members of The National Institute for Japanese Language and are called short-unit word (SUW) and long-unit word (LUW). An SUW approximates a dictionary item found in an ordinary Japanese dictionary, and an LUW represents various compounds. The length and part-of-speech (POS) of each are different, and every SUW is included in an LUW, which is shorter than a Japanese phrasal unit, a bunsetsu. If Manuscript received May 16, 2003; revised October 7, 2003. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Tanja Schultz. K. Uchimoto, C. Nobata, and H. Isahara are with the National Institute of Information and Communications Technology, Soraku-gun, Kyoto 619-0289, Japan (e-mail: uchimoto@crl.go.jp; nova@crl.go.jp; ark@crl.go.jp; isahara@ crl.go.jp). K. Takaoka is with the Justsystem Corporation, Tokushima 771-0189, Japan (e-mail: kazuma-t@is.aist-nara.ac.jp). A. Yamada is with the Advanced Software Technology and Mechatronics Research Institute of Kyoto, Kyoto 600-8813, Japan (e-mail: yamada@ astem.or.jp). S. Sekine is with New York University, New York, NY 10003 USA (e-mail: sekine@cs.nyu.edu). Digital Object Identifier 10.1109/TSA.2004.828700 all of the SUWs in the CSJ were detected, the number of the words would be approximately seven million. That would be the largest spontaneous speech corpus in the world. So far, approximately one tenth of the words have been manually detected, and morphological information such as POS category and conjugation type have been assigned to them. Human annotators tagged every morpheme in the one tenth of the CSJ that has been tagged, and other annotators checked them. The human annotators discussed their disagreements and resolved them. The accuracies of the manual tagging of SUWs and LUWs in the one tenth of the CSJ were greater than 99.8% and 97%, respectively. The accuracies were evaluated by random sampling. As it took over two years to tag one tenth of the CSJ accurately, tagging the remainder with morphological information would take about twenty years. Therefore, the remaining nine tenths of the CSJ must be tagged automatically or semi-automatically. In this paper, we describe methods for detecting the two types of word segments and corresponding morphological information. We also describe how to tag a large spontaneous speech corpus accurately. Henceforth, we call the two types of word segments short-unit word (SUW) and long-unit word (LUW) respectively, or merely morphemes. We use the term morphological analysis for the process of segmenting a given sentence into a row of morphemes and assigning to each morpheme grammatical attributes such as a POS category. II. PROBLEMS AND THEIR SOLUTIONS As we mentioned in Section I, tagging the whole of the CSJ manually would be difficult. Therefore, we are taking a semiautomatic approach. This section describes major problems in tagging a large spontaneous speech corpus with high precision in a semi-automatic way, and our solutions to those problems. One of the most important problems in morphological analysis is that posed by unknown words, which are words found in neither a dictionary nor a training corpus. Two statistical approaches have been applied to this problem. One is to find unknown words from corpora and put them into a dictionary (e.g., [2]), and the other is to estimate a model that can identify unknown words correctly (e.g., [3], [4]). Uchimoto et al. used both approaches. They proposed a morphological analysis method based on a maximum entropy (ME) model [5]. Their method uses a model that estimates how likely a string is to be a morpheme as its probability, and thus it has a potential to overcome the unknown word problem. Therefore, we use their method for morphological analysis of the CSJ. However, Uchimoto et al. reported that the accuracy of automatic word segmentation and 1063-6676/04$20.00 2004 IEEE

UCHIMOTO et al.: MORPHOLOGICAL ANALYSIS OF THE CORPUS OF SPONTANEOUS JAPANESE 383 POS tagging was 94 points in F-measure [6]. That is much lower than the accuracy obtained by manual tagging. Several problems led to this inaccuracy. In the following, we describe these problems and our solutions to them. Filled pauses and disfluencies: Filled pauses and disfluencies are characteristic expressions often used in spoken language, but they are randomly inserted into text, so detecting their segmentation is difficult. In the CSJ, they are tagged manually. Therefore, we first delete filled pauses and word fragments and then put them back in their original places after analyzing a text. Accuracy for unknown words: The morpheme model that will be described in Section III-A can detect word segments and their POS categories even for unknown words. However, the accuracy for unknown words is lower than that for known words. One of the solutions is to use dictionaries developed for a corpus on another domain to reduce the number of unknown words, but the improvement achieved is slight [6]. We believe this is because definitions of a word segment and its POS category depend on a particular corpus, and the definitions from corpus to corpus differ word by word. Therefore, we need to put only words extracted from the same corpus into a dictionary. We are manually examining words that are detected by the morpheme model but that are not found in a dictionary. We are also manually examining those words that the morpheme model estimated as having low probability. During the process of manual examination, if we find words that are not found in a dictionary, those words are then put into a dictionary. Section IV-B1 will describe the accuracy of detecting unknown words and show how much those words contribute to improving the morphological analysis accuracy when they are detected and put into a dictionary. Insufficiency of features: The model currently used for morphological analysis considers the information of a target morpheme and that of an adjacent morpheme on the left. To improve the model, we need to consider the information of two or more morphemes on the left of the target morpheme. However, too much information often leads to overtraining the model. Using all the information makes training the model difficult when there is too much of it. Therefore, the best way to improve the accuracy of the morphological information in the CSJ within the limited time available to us is to examine and revise the errors of automatic morphological analysis as well as to improve the model. We assume that the smaller the probability estimated by a model for an output morpheme is, then the greater the likelihood is that the output morpheme is wrong. Therefore, we examine output morphemes in ascending order of their probabilities. The expected improvement of the accuracy of the morphological information in the whole of the CSJ will be described in Section IV-B1 Another problem concerning unknown words is that the cost of manual examination is high when there are several definitions for word segments and their POS categories. Since there are two types of word definitions in the CSJ, Fig. 1. Example of transcription. the cost would double. Therefore, to reduce the cost, we propose another method for detecting word segments and their POS categories. The method will be described in Section III-B, and the advantages of the method will be described in Section IV-B2 The next problem described here is one that we have to solve to make a language model for automatic speech recognition. Phonetic transcription: Phonetic transcription of each word is indispensable for making a language model for automatic speech recognition. In the CSJ, pronunciation is transcribed separately from the orthographic transcription written by using kanji and hiragana characters as shown in Fig. 1. Text targeted for morphological analysis is the orthographic transcription of the CSJ and it does not have information on actual pronunciation. The result of morphological analysis, therefore, is a row of morphemes that do not have information on actual pronunciation. To estimate actual pronunciation by using only the orthographic transcription and a dictionary is impossible. Therefore, actual pronunciation is assigned to results of morphological analysis by aligning the orthographic transcription and phonetic transcription in the CSJ. First, the results of morphological analysis, namely, the morphemes, are transliterated into katakana characters by using a dictionary, and then they are aligned with phonetic transcription in the CSJ by using a dynamic programming method. In this paper, we will mainly discuss methods for detecting word segments and their POS categories in the whole of the CSJ. III. MODELS AND ALGORITHMS This section describes two methods for detecting word segments and their POS categories. The first method uses morpheme models and is used to detect any type of word segment. The second method uses a chunking model and is only used to detect LUW segments. A. Morpheme Model Given a tokenized test corpus, namely a set of strings, the problem of Japanese morphological analysis can be reduced to the problem of assigning one of two tags to each string in a sentence. A string is tagged with a 1 or a 0 to indicate whether it is a morpheme. When a string is a morpheme, a grammatical attribute is assigned to it. A tag designated as a 1 is thus assigned one of a number,, of grammatical attributes assigned to morphemes, and the problem becomes to assign an attribute (from 0 to ) to every string in a given sentence. We define a model that

384 IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 4, JULY 2004 Fig. 2. Example of morphological analysis results. estimates the likelihood that a given string is a morpheme and has a grammatical attribute as a morpheme model. We implemented this model within an ME modeling framework [7] [9]. The model is represented by where is one of the categories for classification, and it can be one of tags from 0 to (called a future), is the contextual or conditioning information that enables us to make a decision among the space of futures (called a history), and is a normalizing constant determined by the requirement that for all. The computation of in any ME model is dependent on a set of features which are binary functions of the history and future. For instance, one of our features is Here, is a binary function that returns 1 if the history has feature. The features used in our experiments are described in detail in Section IV-A1. Given a sentence, probabilities of tags from 1 to are estimated for each length of string in that sentence by using the morpheme model. From all possible division of morphemes in the sentence, an optimal one is found by using the Viterbi algorithm. Each division is represented as a particular division of morphemes with grammatical attributes in a sentence, and the optimal division is defined as a division that maximizes the product of the probabilities estimated for each morpheme in the division. For example, the sentence (1) (2) (3) in orthographic transcription as shown in Fig. 1 is analyzed as shown in Fig. 2. is analyzed as three morphemes, (noun), (suffix), and (noun), for SUWs, and as one morpheme, (noun), for LUWs. B. Chunking Model The model described in this section can be applied when several types of words are defined in a corpus and one type of words consists of compounds of other types of words. In the CSJ, every LUW consists of one or more SUWs. Our method uses two models, a morpheme model for SUWs and a chunking model for LUWs. After detecting SUW segments and their POS categories by using the former model, LUW segments and their POS categories are detected by using the latter model. We define four labels, as explained below, and extract LUW segments by estimating the appropriate labels for each SUW according to an ME model. The four labels are listed as follows: Ba beginning of an LUW, and the POS category of the LUW agrees with the SUW; Ia middle or end of an LUW, and the POS category of the LUW agrees with the SUW; B beginning of an LUW, and the POS category of the LUW does not agree with the SUW; I middle or end of an LUW, and the POS category of the LUW does not agree with the SUW. A label assigned to the leftmost constituent of an LUW is Ba or B. Labels assigned to other constituents of an LUW are Ia or I. For example, the SUWs shown in Fig. 2 are labeled as shown in Fig. 3. The labeling is done deterministically from the beginning of a given sentence to its end. The label that has the highest probability as estimated by an ME model is assigned to each SUW. The model is represented by (1). In (1), can be one of four labels. The features used in our experiments are described in Section IV-A2. When an LUW that does not include an SUW that has been assigned the label Ba or Ia, this indicates that the word s POS category differs from all of the SUWs that constitute the LUW. Such a word must be estimated individually. In this case, we estimate the POS category by using transformation rules. The

UCHIMOTO et al.: MORPHOLOGICAL ANALYSIS OF THE CORPUS OF SPONTANEOUS JAPANESE 385 Fig. 3. Example of labeling. transformation rules are automatically acquired from the training corpus by extracting LUWs with constituents, namely SUWs, that are labeled only B or I. A rule is constructed by using the extracted LUW and the adjacent SUWs on its left and right. For example, the rule shown in Fig. 4 was acquired in our experiments. The middle division of the consequent part represents an LUW (auxiliary verb), and it consists of two SUWs (post-positional particle) and (verb). If several different rules have the same antecedent part, only the rule with the highest frequency is chosen. If no rules can be applied to an LUW segment, rules are generalized in the following steps. 1) Delete posterior context. 2) Delete anterior and posterior contexts. 3) Delete anterior and posterior contexts and lexical entries. If no rules can be applied to an LUW segment in any step, the POS category noun is assigned to the LUW. IV. EXPERIMENTS AND DISCUSSION A. Experimental Conditions In our experiments, we used 744 204 SUWs and 618 538 LUWs for training, and 63 037 SUWs and 51 796 LUWs for testing. Those words were extracted from one tenth of the CSJ that already had been manually tagged. The training corpus consisted of 319 talks and the test corpus consisted of 19 talks. Transcription consisted of orthographic transcription and phonetic transcription, as shown in Fig. 1. Lecture speeches were faithfully transcribed as phonetic transcription, and also represented as orthographic transcriptions by using kanji and hiragana characters. Lines beginning with numerical digits are time stamps and represent the time it took to produce the lines between that time stamp and the next time stamp. Each line other than time stamps represents a bunsetsu. In our experiments, we used only the orthographic transcriptions. Orthographic transcriptions were tagged with several types of labels such as filled pauses, as shown in Table I. Strings tagged with those labels were handled according to rules as shown in the rightmost columns in Table I. Since there are no boundaries between sentences in the corpus, we selected the places in the CSJ that are automatically detected as pauses of 500 ms or longer and then designated them as sentence boundaries. In addition to these, we also used utterance boundaries with short pauses as sentence boundaries. These are automatically detected at places where short pauses (shorter than 200 ms but longer than 50 ms) follow the typical sentence-ending forms of predicates such as verbs, adjectives, and copula. 1) Features Used by Morpheme Models: In the CSJ, bunsetsu boundaries, which are phrase boundaries in Japanese, were manually detected. Filled pauses and word fragments were marked with the labels (F) and (D). In the experiments, we eliminated filled pauses and word fragments but we did use their positional information as features. We also used as features, bunsetsu boundaries and the labels (M), (O), (R), and (A), which were assigned to particular morphemes such as personal names and foreign words. Thus, the input sentences for training and testing were character strings without filled pauses and word fragments, and both boundary information and various labels were attached to them. Given a sentence, for every string within a bunsetsu and every string appearing in a dictionary, the probabilities of in (1) were estimated by using the morpheme model. The output was a sequence of morphemes with grammatical attributes, as shown in Fig. 2. We used the POS categories in the CSJ as grammatical attributes. We obtained 14 major POS categories for SUWs and 15 major POS categories for LUWs. Therefore, in (1) can be one of 15 tags from 0 to 14 for SUWs, and it can be one of 16 tags from 0 to 15 for LUWs. The features we used with morpheme models in our experiments are listed in Table II. Each feature consists of a type and a value, which are given in the rows of the table, and it corresponds to in the function in (1). The notations (0) and used in the feature-type column in Table II, respectively indicate a target string and the morpheme to the left of it. The terms used in the table are basically the same as those used by Uchimoto et al. [6]. The main difference is the following. Boundary: Bunsetsu boundaries and positional information of labels such as filled pauses. (Beginning) and (End) in Table II, respectively, indicate whether the left and right side of the target strings are boundaries. We used only those features that were found three or more times in the training corpus. 2) Features Used by a Chunking Model: We used the following information as features on the target word: a word and its POS category, and the same information for the four closest words, the two on the left and the two on the right of the target word. Bigram and trigram words that included a target word plus bigram and trigram POS categories that included the target word s POS category were used as features. In addition, bunsetsu boundaries as described in Section IV-A.1 were used. For example, when a target word was in Fig. 3,,,,,, Suffix, Noun, PPP, Verb, PPP,,,,, Noun&PPP, PPP&Verb, Suffix&Noun&PPP, PPP&Verb&PPP, and Bunsetsu(Beginning) were used as features. B. Results and Discussion 1) Experiments Using Morpheme Models: Results of the morphological analysis obtained by using morpheme models

386 IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 4, JULY 2004 Fig. 4. Example of transformation rules. TABLE I TYPE OF LABELS AND THEIR HANDLING TABLE II FEATURES are shown in Tables III and IV. In these tables, out-of-vocabulary rates are indicated as OOV. As shown in Table III, OOV was calculated as the proportion of words not found in a dictionary to all words in the test corpus. In Table IV, OOV was calculated as the proportion of word and POS category pairs that were not found in a dictionary to all pairs in the test corpus. Recall is the percentage of morphemes in the test corpus for which the segmentation and major POS category were identified correctly. Precision is the percentage of all morphemes identified by the system that were identified correctly. The F-measure is defined by the following equation: Tables III and IV show that accuracies would improve significantly if no words were unknown. This indicates that all morphemes of the CSJ could be analyzed accurately if there were no unknown words. The improvements that we can expect by

UCHIMOTO et al.: MORPHOLOGICAL ANALYSIS OF THE CORPUS OF SPONTANEOUS JAPANESE 387 TABLE III ACCURACIES OF WORD SEGMENTATION TABLE IV ACCURACIES OF WORD SEGMENTATION AND POS TAGGING detecting unknown words and putting them into dictionaries are about 1.5 in F-measure for detecting word segments of SUWs and 2.5 for LUWs. For detecting the word segments and their POS categories, for SUWs we expect an improvement of about 2 in F-measure and 3 for LUWs. Next, we discuss accuracies obtained when unknown words existed. The OOV for LUWs was 4% higher than that for SUWs. In general, the higher the OOV is, the more difficult detecting word segments and their POS categories is. However, the difference between accuracies for SUWs and LUWs was about 1% in recall and 2% in precision, which is not significant when we consider that the difference between OOVs for SUWs and LUWs was 4%. This result indicates that our morpheme models could detect both known and unknown words accurately, especially LUWs. Therefore, we investigated the recall of unknown words in the test corpus, and found that 55.7% (928/1667) of SUW segments and 74.1% (2660/3590) of LUW segments were detected correctly. In addition, regarding unknown words, we also found that 47.5% (791/1667) of SUW segments plus their POS categories and 67.3% (2415/3590) of LUW segments plus their POS categories were detected correctly. The recall of unknown words was about 20% higher for LUWs than for SUWs. We believe that this result mainly depended on the difference between SUWs and LUWs in terms of the definitions of compound words. A compound word is defined as one word when it is based on the definition of LUWs; however it is defined as two or more words when it is based on the definition of SUWs. Furthermore, based on the definition of SUWs, a division of compound words depends on its context. More information is needed to precisely detect SUWs than is required for LUWs. Next, we extracted words that were detected by the morpheme model but were not found in a dictionary, and investigated the percentage of unknown words that were completely or partially matched to the extracted words by their context. This percentage was 77.6% (1293/1667) for SUWs, and 80.6% (2892/3590) for LUWs. Most of the remaining unknown words that could not be detected by this method are compound words. We expect that these compounds can be detected during the manual examination of those words for which the morpheme model estimated a low probability, as will be shown later. The recall of unknown words was lower than that of known words, and the accuracy of automatic morphological analysis was lower than that of manual morphological analysis. As previously stated, to improve the accuracy of the whole corpus we take a semi-automatic approach. We assume that the smaller the probability is for an output morpheme estimated by a model, the more likely the output morpheme is wrong, and we examine output morphemes in ascending order of their probabilities. We investigated how much the accuracy of the whole corpus would increase. Fig. 5 shows the relationship between the percentage of output morphemes whose probabilities exceed a threshold and their precision. In this figure, short without UKW, long without UKW, short with UKW, and long with UKW represent the precision for SUWs detected assuming there were no unknown words, precision for LUWs detected assuming there were no unknown words, precision of SUWs including unknown words, and precision of LUWs including unknown words, respectively. When the output rate in the horizontal axis increases, the number of low-probability morphemes increases. In all graph lines, precisions monotonously decrease as output rates increase. This means that tagging errors can be revised effectively when morphemes are examined in ascending order of their probabilities. Next, we investigated the relationship between the percentage of morphemes examined manually and the precision obtained after detected errors were revised. The result is shown in Fig. 6. Precision represents the precision of word segmentation and POS tagging. If unknown words were detected and put into a dictionary by the method described earlier in this section, the graph line for SUWs would be drawn between the graph lines short without UKW and short with UKW, and the graph line for LUWs would be drawn between the graph lines long without UKW and long with UKW. Based on test results, we can expect better than 99% precision for SUWs and better than 97% precision for LUWs in the whole corpus when we examine 10% of output morphemes in ascending order of their probabilities. Finally, we investigated the relationship between percentage of morphemes examined manually and the error rate for all of the examined morphemes. The result is shown in Fig. 7. We found that about 50% of examined morphemes would be found as errors at the beginning of the examination and about 20% of examined morphemes would be found as errors when examination of 10% of the whole corpus was completed. When unknown words were detected and put into a dictionary, the error rate decreased; even so, over 10% of examined morphemes would be found as errors. 2) Experiments Using Chunking Models: Results of the morphological analysis of LUWs obtained by using a chunking model are shown in Tables V and VI. The first and second lines show the respective accuracies obtained when OOVs were 5.81% and 6.93%. The third lines show the accuracies obtained when we assumed that the OOV for SUWs was 0% and there were no errors in detecting SUW segments and their POS categories. The accuracy obtained by using the chunking model was one point higher in F-measure than that obtained by using the morpheme model, and it was very close to the accuracy achieved for

388 IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 4, JULY 2004 Fig. 5. Partial analysis. Fig. 6. Relationship between the percentage of morphemes examined manually and precision obtained after revising detected errors (when morphemes with probabilities under threshold and their adjacent morphemes are examined). SUWs. This result indicates that errors newly produced by applying a chunking model to the results obtained for SUWs were slight, or errors in the results obtained for SUWs were amended by applying the chunking model. This result also shows that we can achieve good accuracy for LUWs by applying a chunking model even if we do not detect unknown LUWs and do not put them into a dictionary. If we could improve the accuracy for SUWs, the accuracy for LUWs would be improved also. The third lines in Tables V and VI show that the accuracy would improve to over 98 points in F-measure. Considering the results obtained in this section and in Section IV-B1, we are now detecting SUW and LUW segments and their POS categories in the whole corpus by using the following steps: 1) Automatically detect and manually examine unknown words for SUWs.

UCHIMOTO et al.: MORPHOLOGICAL ANALYSIS OF THE CORPUS OF SPONTANEOUS JAPANESE 389 Fig. 7. Relationship between percentage of morphemes examined manually and error rate of the examined morphemes. TABLE V ACCURACIES OF LUW SEGMENTATION TABLE VI ACCURACIES OF LUW SEGMENTATION AND POS TAGGING The training corpus consisted of 396 talks. 1. The number of SUWs detected by our system was 7 004 124, and after putting back the deleted filled pauses and word fragments in their original places, the number of SUWs was 7 591 683. 20 000 SUWs were randomly sampled from the 7 591 683 SUWs, and manually evaluated. The precision of SUW segmentation and POS information tagging was 97.16% (19 431/20 000), and the precision of SUW without filled pauses and word fragments was 97.00% (17 869/18 421). It is close to the precision obtained when we assumed that the OOV for SUWs was 0%. The result shows the effectiveness of the first step of our approach. Other steps are ongoing. 2) Manually examine SUWs extracted by applying an active learning method proposed by Argamon-Engelson and Dagan [10], and use the examined SUWs for training a morpheme model. 3) Improve the accuracy for SUWs in the whole corpus by manually examining SUWs in ascending order of their probabilities estimated by a morpheme model. 4) Apply a chunking model to the SUWs to detect LUW segments and their POS categories. After the first step, we analyzed the nine-tenths of the CSJ that were not tagged manually, and evaluated the accuracy for SUWs by random sampling. The number of the extracted and manually examined unknown words for SUWs was 67 574. We used 909 468 SUWs for training that were extracted from the one tenth of the CSJ that already had been manually tagged. V. CONCLUSION This paper described two methods for detecting word segments and their POS categories in a Japanese spontaneous speech corpus, and describes how to tag a large spontaneous speech corpus accurately by using the two methods. The first method is used to detect any type of word segment. We found that about 80% of unknown words could be semi-automatically detected by using this method. The second method is used when there are several definitions for word segments and their POS categories, and when one type of word segments includes another type of word segments. We found that better accuracy could be achieved by using both methods than by using only the first method alone. Two types of word segments, SUWs and LUWs, are found in a large spontaneous speech corpus, CSJ. We found that the ac- 1 The number of manually tagged talks increased after the experiments as shown in Section IV-B1

390 IEEE TRANSACTIONS ON SPEECH AND AUDIO PROCESSING, VOL. 12, NO. 4, JULY 2004 curacy of automatic morphological analysis for the SUWs was 95.79 in F-measure and for LUWs, 95.49. Although the OOV for LUWs was much higher than that for SUWs, almost the same accuracy was achieved for both types of words by using our proposed methods. We also found that we can expect more than 99% of precision for SUWs, and 97% for LUWs found in the whole corpus when we examined 10% of output morphemes in ascending order of their probabilities as estimated by the proposed models. Kazuma Takaoka received the M.S. degree in information science from Nara Institute of Science, Nara, Japan, in 2000. He is a Researcher with the Justsystem Corporation, Tokushima, Japan. He was previously with Nara Institute of Science, Nara, where this work was performed. His current research interest includes Japanese morphological analysis and a syntactic and semantic representation of compound verbs. ACKNOWLEDGMENT The authors would like to thank Prof. S. Furui of Tokyo Institute of Technology for his supervision of the Spontaneous Speech: Corpus and Processing Technology project, and the members of The National Institute for Japanese Language who were involved in the project, especially Dr. M. Yamaguchi, Dr. H. Ogura, Dr. K. Nishikawa, Dr. H. Koiso, and Dr. K. Maekawa for their advice and evaluation of our system. Chikashi Nobata received the Ph.D. degree from the Department of Computer Science, University of Tokyo, Tokyo, Japan, in 2000. He joined the Computational Linguistics Group at the National Institute of Information and Communications Technology, Kyoto, Japan, in 2000. His current research interests include information extraction and automatic summarization. REFERENCES [1] K. Maekawa, H. Koiso, S. Furui, and H. Isahara, Spontaneous speech corpus of Japanese, in Proc. LREC2000, 2000, pp. 947 952. [2] S. Mori and M. Nagao, Word extraction from corpora and its part-ofspeech estimation using distributional analysis, in Proc. 16th Int. Conf. Computational Linguistics (COLING96), 1996, pp. 1119 1122. [3] H. Kashioka, S. G. Eubank, and E. W. Black, Decision-tree morphological analysis without a dictionary for Japanese, in Proc. Natural Language Processing Pacific Rim Symp., 1997, pp. 541 544. [4] M. Nagata, A part of speech estimation method for Japanese unknown words using a statistical model of morphology and context, in Proc. 37th Annu. Meeting Assoc. Computational Linguistics (ACL), 1999, pp. 277 284. [5] K. Uchimoto, S. Sekine, and H. Isahara, The unknown word problem: A morphological analysis of Japanese using maximum entropy aided by a dictionary, in Proc. 2001 Conf. Empirical Methods in Natural Language Processing, 2001, pp. 91 99. [6] K. Uchimoto, C. Nobata, A. Yamada, S. Sekine, and H. Isahara, Morphological analysis of the spontaneous speech corpus, in Proc. 19th Int. Conf. Computational Linguistics (COLING2002), 2002, pp. 1298 1302. [7] E. T. Jaynes, Information theory and statistical mechanics, Phys. Rev., vol. 106, pp. 620 630, 1957. [8], Where do we stand on maximum entropy?, in The Maximum Entropy Formalism, R. D. Levine and M. Tribus, Eds. Cambridge, MA: MIT Press, 1979. [9] A. L. Berger, S. A. D. Pietra, and V. J. D. Pietra, A maximum entropy approach to natural language processing, Comput. Linguist., vol. 22, no. 1, pp. 39 71, 1996. [10] S. Argamon-Engelson and I. Dagan, Committee-based sample selection for probabilistic classifiers, Artif. Intell. Res., vol. 11, pp. 335 360, 1999. Kiyotaka Uchimoto received the B.E. and M.E. degrees in electrical engineering, and the Ph.D. degree in informatics, from Kyoto University, Kyoto, Japan, in 1994, 1996, and 2004, respectively. He is a Senior Research Scientist of the National Institute of Information and Communications Technology, Kyoto, Japan. His main research area is corpus-based natural language processing, and he specializes in Japanese sentence analysis and generation and information extraction. Dr. Uchimoto is a member of the Association for Natural Language Processing, the Information Processing Society of Japan, and the Association for Computational Linguistics. Atsushi Yamada received the Ph.D. degree in information science from Kyoto University, Kyoto, Japan, in 1993. He is a Senior Researcher with the Advanced Software Technology and Mechatronics Research Institute, Kyoto. He was previously an Expert Researcher with the National Institute of Information and Communications Technology, Kyoto. His main research area is language processing, including text processing for spoken language, text annotation, data or meta-data conversion and transformation, and other XML-related matters. Satoshi Sekine received the Ph.D. degree from the Computer Science Department, New York University (NYU), in 1998. He is an Assistant Research Professor at the Computer Science Department, NYU. His main research interests are in the area of natural language processing, including English and Japanese sentence analyzers, sublanguage study, information extraction, question answering, and summarization. Hitoshi Isahara received the B.E., M.E., and Ph.D. degrees in electrical engineering from Kyoto University, Kyoto, Japan, in 1978, 1980, and 1995, respectively. He is a Leader of the Computational Linguistics Group at the National Institute of Information and Communications Technology, Kyoto, Japan. He is also a Professor at Kobe University Graduate School of Science and Technology, Kobe, Japan. His research interests include natural language processing, machine translation, and lexical semantics. Dr. Isahara is a member of the Association for Natural Language Processing, the Information Processing Society of Japan, and the Association for Computational Linguistics.